Research Decision

Iteration Decision Support

Analyze the state of a completed hypothesis experiment and the overall hypothesis graph to recommend the next action.

Context

After each experiment, the researcher makes one of four decisions:

• Continue — hypothesis confirmed, deepen this direction, formulate

child hypotheses.

• Pivot — result points to a more promising direction different from

the current one. Formulate new hypotheses, change vector.

• Kill — branch is a dead end (refuted, unstable, impractical). Prune

the branch in the graph, document the reason.

• Fork — competing approaches discovered that are worth investigating

in parallel. Create multiple child hypotheses for simultaneous exploration.

The decision is the researcher's, but this skill provides structured analysis and a recommendation to support it.

Procedure

Step 1 — Read current state

Gather the following information:

1. The completed hypothesis card — read hypotheses/H-NNN.md for the

statement, verification criterion, result, and status.

1. The hypothesis graph — read hypotheses/graph.md to understand the

full DAG: which branches are alive, which are pruned, where pivots occurred, and what the current "front line" looks like.

1. The research brief — read brief.md for the research question,

success criteria, scope boundaries, known constraints, and time budget.

1. Sibling/related hypotheses — if the completed hypothesis is part of

a fork, read its sibling hypotheses to compare results.

Step 2 — Analyze the result

Evaluate the experiment outcome against multiple dimensions:

2a. Verification criterion alignment

• Was the verification criterion met, partially met, or not met?
• Is the result definitive or are there caveats?
• If inconclusive: is the inconclusiveness due to the approach, the

timebox, or an external factor?

2b. Research question proximity

• How much closer does this result bring us to answering the research

question from the brief?

• Does the result open new pathways toward the answer?
• Does the result suggest the research question itself needs refinement?

2c. Resource and timebox assessment

• How much of the overall research time budget has been consumed?
• How many active branches remain in the graph?
• Is there sufficient budget to pursue new directions, or should the

research begin converging?

2d. Surprise and serendipity

• Did the experiment reveal unexpected findings outside the original scope?
• Are there serendipitous discoveries worth pursuing (potential fork)?

Step 3 — Generate recommendation

Based on the analysis, recommend one of the four decisions with justification:

Continue — recommend when:

• The hypothesis was confirmed.
• The result directly advances the research question.
• There are clear, specific follow-up questions.
• Budget allows further depth.

When recommending Continue, propose 1–3 specific child hypotheses with draft statements and verification criteria.

Pivot — recommend when:

• The result (positive or negative) reveals a more promising direction.
• The current branch has diminishing returns.
• A related but different approach looks more viable.

When recommending Pivot, clearly articulate:

• What the new direction is and why it is more promising.
• What is being left behind and why.
• Proposed hypothesis(es) for the new direction.

Kill — recommend when:

• The hypothesis was definitively refuted.
• The timebox expired without meaningful progress.
• The branch is technically viable but impractical (too complex, too slow,

too fragile for the intended use case).

• Continuing this branch will not help answer the research question.

When recommending Kill, document:

• The specific reason for termination.
• Any useful knowledge gained that should be preserved.
• Whether this dead end has implications for other branches.

Fork — recommend when:

• Two or more competing approaches are identified.
• It is unclear which approach is superior without trying both.
• Resources are available for parallel investigation.

When recommending Fork, propose:

• The competing hypotheses (2–3 maximum).
• Individual timeboxes for each.
• Selection criteria: how will we choose the winner (or discard both)?

Step 4 — Present to the researcher

Present the analysis and recommendation in a structured format:

1. Result summary: one paragraph on what the experiment showed.
2. Graph context: where this result fits in the overall research.
3. Recommendation: the decision (Continue / Pivot / Kill / Fork) with

reasoning.

1. Proposed next hypotheses: if applicable, draft cards for the next

hypotheses.

1. Risks and alternatives: what could go wrong with the recommendation,

and what the alternative decision would be.

Step 5 — Execute the decision

Once the researcher confirms (or modifies) the decision:

1. Update the completed hypothesis card with the decision in the

"Decision" field.

1. If the decision is Continue or Fork: invoke research-hypothesis

to create the new child hypothesis cards.

1. If the decision is Pivot: invoke research-hypothesis to create new

hypotheses. Mark the abandoned branch hypotheses as cancelled if they are still open.

1. If the decision is Kill: ensure the hypothesis status is refuted

or cancelled as appropriate. No new hypotheses are created on this branch.

Decision Heuristics

Some rules of thumb encoded in the methodology:

• Timebox expiry + inconclusive = kill or pivot, not extension. The

methodology explicitly discourages extending timeboxes.

• Budget pressure → convergence: when more than ~70% of the overall

research budget is consumed, prefer Continue (deepening) over Fork (broadening). Begin thinking about synthesis.

• Fork is the exception: parallel investigation is allowed but should

be justified. The default is sequential hypothesis testing.

• Preserve knowledge from dead ends: killed branches still produced

knowledge. Ensure the "what we learned" section of the card is thorough.

Relation to Other Skills

• Invoked after research-experiment records a result.
• Creates new hypotheses via research-hypothesis.
• Feeds into research-synthesis when the research question is answered

or budget is exhausted (triggers Phase 3).

• research-status provides the graph overview that informs decisions.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: v0lka
• Source: v0lka/skills
• License: MIT

Install and usage instructions live in the source repository linked above.